On Out-of-Domain Generalization in Semi-Supervised Pedestrian Attribute Recognition

カテゴリ: 研究会(論文単位)

論文No: CMN24025

グループ名: 【C】電子・情報・システム部門 通信研究会

発行日: 2024/03/25

タイトル(英語): On Out-of-Domain Generalization in Semi-Supervised Pedestrian Attribute Recognition

著者名: Ravikiran Manikandan(Hitachi India Pvt Ltd),Kumar Sharath(Hitachi India Pvt Ltd),Ganesh Ananth(Hitachi India Pvt Ltd)

著者名(英語): Manikandan Ravikiran(Hitachi India Pvt Ltd),Sharath Kumar(Hitachi India Pvt Ltd),Ananth Ganesh(Hitachi India Pvt Ltd)

キーワード: Deep Learning, |Pedestrian Attribute Recognition|Semi-supervised Learning|Deep Learning, |Pedestrian Attribute Recognition|Semi-supervised Learning

要約(日本語): Semi Supervised Pedestrian attribute recognition (Semi-PAR) focuses on identification of attributes of people given their input image, with less annotated samples. Recent works have shown Hierarchical Groupwise Temporal Ensemble (Hi-GOTE) based Semi-PAR improve performance over other methods by training groupwise encoders, however there is currently lack of comprehensive evaluation on out-of-domain datasets. Accordingly in this paper we empirically answer (a) How does Hi-GOTE fare against strong out-of-domain data (b) How effective Hi-GOTE is on coarse grained and fine-grained attributes (c) How does Hi-GOTE's generalization performance vary with addition of labelled samples from out-of-domain domain dataset. Empirical results against a novel out-of-domain dataset Hi-ODATA reveals that (a) Hi-GOTE's performance is on par for strong out-of-domain data (b) it performs well against coarse grained attributes and (c) addition of out-of-domain samples improves performance of Hi-GOTE by additional ~2% in accuracy.

要約(英語): Semi Supervised Pedestrian attribute recognition (Semi-PAR) focuses on identification of attributes of people given their input image, with less annotated samples. Recent works have shown Hierarchical Groupwise Temporal Ensemble (Hi-GOTE) based Semi-PAR improve performance over other methods by training groupwise encoders, however there is currently lack of comprehensive evaluation on out-of-domain datasets. Accordingly in this paper we empirically answer (a) How does Hi-GOTE fare against strong out-of-domain data (b) How effective Hi-GOTE is on coarse grained and fine-grained attributes (c) How does Hi-GOTE's generalization performance vary with addition of labelled samples from out-of-domain domain dataset. Empirical results against a novel out-of-domain dataset Hi-ODATA reveals that (a) Hi-GOTE's performance is on par for strong out-of-domain data (b) it performs well against coarse grained attributes and (c) addition of out-of-domain samples improves performance of Hi-GOTE by additional ~2% in accuracy.

本誌: 2024年3月28日-2024年3月29日通信研究会

本誌掲載ページ: 43-47 p

原稿種別: 英語

PDFファイルサイズ: 878 Kバイト